When Serverless Computing Meets Different Degrees of Customization for DNN Inference

Serverless computing provides a method to develop application services without the burden of run-time execution environment management overhead. Since the initial offerings of serverless computing using function-as-a-service (FaaS), other variants of execution environments have been proposed, such as a special-purpose FaaS (SPF) for deep neural network (DNN) inference and a serverless container service (SCS) for general web applications. This paper qualitatively summarizes the characteristics of a general-purpose FaaS (GPF), SPF, and SCS from the perspective of customizability when setting up execution environments. To judge whether various serverless computing environments can be feasible solutions for an interactive DNN model inference application, we conduct extensive experiments and conclude that there are rooms for performance improvement serverless DNN inference, and allowing a custom environment setup can make the serverless computing platform for an interactive DNN application.